Programmers refuse to work without AI, and this may affect them

Researchers have discovered that in 2026, you can’t wrest AI programming tools from developers.

But while AI undoubtedly helps programmers produce code faster, it may not produce better code, other researchers warn. This could cause problems for them in the future.

Specifically, in February 2026, the respected AI research lab METR published a surprising revelation: most developers, even on a limited number of tasks, will no longer work without AI.

METR was hoping to provide an update on some groundbreaking research published a few months ago, in 2025, on AI coding productivity. In this study, researchers measured the amount of time it takes open source developers to do tasks manually versus the time it takes artificial intelligence.

While the developers in that study said that AI made them more productive, they were shocked to learn that it actually slowed them down. Sure, they generated code faster, but they spent extra time finding and fixing bugs, directing the AI ​​and waiting for it to complete tasks.

When METR set out to replicate the experiment to measure progress in artificial intelligence and programming efficiency, it couldn’t.

The researchers admitted that developers were not willing to participate “because they did not want to work without the AI” even just for the sake of study.

Instead, METR published a survey in May that allowed technical employees to self-report their AI productivity gains. Not surprisingly, they realized that AI made them doubly valuable to their organizations.

But recent headlines about the enormous cost of so-called Tokenmaxxing, coupled with a smattering of recent research, make such subjective perceptions questionable.

Tokenmaxxing, or using the number of tokens a person uses as a proxy for productivity using AI, has been the trend in 2026 so far. It may already be over.

The Financial Times reported this week that Amazon shut down its internal leaderboard for tracking tokens called Kirorank after employees were manipulating it with excessive use of artificial intelligence agents, which drove up costs. Employees have proven that using AI does not automatically lead to increased productivity.

Uber exceeded its 2026 AI budget during the first four months of the year, The Information reported. Such spending has not led to a measurable increase in projects or productivity, COO Andrew McDonald recently said on a podcast.

And AI-generated code does not necessarily reduce, and may even increase, ongoing code maintenance needs, as programmer and author James Schur elegantly argued in a blog post that went viral on Hacker News.

“Are you writing code twice as fast now? You might as well have cut your maintenance costs in half,” he wrote. “Otherwise you’re in a bad situation. You’re trading a temporary increase in speed for a permanent contract.”

There is other evidence that AI can increase code maintenance problems.

A viral tweet from Aiswarya Sankar, founder and CEO of reliability engineering agent startup Entelligence AI, announces that companies are spending 44% of their tokens on fixes for bugs generated by their AI. Meanwhile, code review tool company CodeRabbit says it analyzed open source pull requests and found that AI generated 1.7 times more issues than human code.

Admittedly, these are self-serving statistics from those trying to sell AI code review tools.

However, independent researchers have also found such problems. Researchers from the prestigious Singapore Management University published a report in April warning that “AI-generated code can lead to long-term maintenance costs in real software projects.”

Since programmers love AI assistants, what’s the solution?

Well, those who want to sell AI coding agents say that developers can only use AI coding agents to do the cumbersome tasks of fixing code as quickly as the AI ​​deploys it. That’s what Cognition founder and CEO Scott Wu, maker of the AI ​​coding agent Devin, suggests.

But even he admits that while Devin can work independently, he currently rates his skill between beginner and intermediate programmers, depending on the task. This is not a just hand it over and forget about it solution.

Researchers at SMU suggest a more humane approach. Programmers should know what AI does and doesn’t do as well as they know their favorite programming languages. They need robust quality assurance systems designed for AI, and they are stuck carefully reviewing the AI’s work as if it were a junior developer.

Meanwhile, the researchers say (and Wu agrees), humans should keep doing the big work like software engineering and security design.